Niklas Wahl

Objective: Deep learning (DL) methods enable photon dose calculation under two main coordinate representations: Beam’s Eye View (BEV) and patient coordinates. We evaluate dose calculation accuracy and speed under these coordinate paradigms and with representative DL models within a unified dataset and pipeline, and introduce two lightweight models for fast photon dose calculation. Approach: Planning computed tomography (CT) scans and volumetric modulated arc therapy (VMAT) plans from 24 prostate cancer patients were used. Monte Carlo (MC) simulation generated 5940, 540, and 3053 segment doses for training (11 patients), validation (3), and testing (10), respectively. For BEV, we used a combination of convolutional neural network (CNN) and convolutional long short-term memory network (ConvLSTM) called CNN-ConvLSTM, a CNN-Mamba combination (CNN-Mamba), a transformer-based architecture (DoTA), and a cascaded 3D UNet (C3D). These were trained on CT and segment-projection BEV cuboids. For patient coordinates, the DeepDose individual segment dose prediction framework implemented with C3D (DeepDose-C3D) was trained on cropped CT volumes with four physical inputs. Segment and plan dose accuracy were assessed using local gamma passing rates γPR (2%/3 mm and 1%/3 mm) and dose–volume histogram metrics. Dose calculation times (inference plus pre/post-processing) were measured on three different GPUs. Results: All five models achieved mean local γPR values ≥91.0% (2%/3 mm) for segment doses and ≥99.0% (1%/3 mm) for plan doses. Mean per-segment dose calculation times were 79, 67, 298, 490, 356 ms for CNN-ConvLSTM, CNN-Mamba, DoTA, C3D, and DeepDose-C3D, respectively. On the latest-generation GPU avaliable, the corresponding per-plan (average 305 segments) dose calculation times were 5.5, 6.2, 33.6, 38.7 35.4 s. Significance: Both BEV- and patient-coordinate DL methods achieved accurate photon plan dose calculation, with BEV-based approaches showing more robust segment performance. CNN-ConvLSTM and CNN-Mamba retain comparable accuracy at lower computational cost, enabling fast photon dose calculation.

Background: In proton therapy of low‐grade glioma (LGG) patients, contrast‐enhancing brain lesions (CEBLs) on magnetic resonance imaging are considered predictive of late radiation‐induced lesions. From the observation that CEBLs tend to concentrate in regions of increased dose‐averaged linear energy transfer (LET _d ) and proximal to the ventricular system, the probability of lesion origin (POLO) model has been established as a multivariate logistic regression model for the voxel‐wise probability prediction of the CEBL origin. Purpose: To date, leveraging the predictive power of the POLO model for treatment planning relies on hand tuning the dose and LET _d distribution to minimize the resulting probability predictions. In this paper, we therefore propose automated POLO model‐based treatment planning by directly integrating POLO calculation and optimization into plan optimization for LGG patients. Approach: We introduce an extension of the original POLO model including a volumetric correction factor, and a model‐based optimization scheme featuring a linear reformulation of the model together with feasible optimization functions based on the predicted POLO values. The developed framework is implemented in the open‐source treatment planning toolkit matRad. Results: Our framework can generate clinically acceptable treatment plans while automatically taking into account outcome predictions from the POLO model. It also supports the definition of customized POLO model‐based objective and constraint functions. Optimization results from a sample LGG patient show that the POLO model‐based outcome predictions can be minimized under expectable shifts in dose, LET _d , and POLO distributions, while sustaining target coverage (, ), even when NTCP is strongly downregulated. Conclusion: POLO model‐based treatment plan optimization for LGG patients can be implemented in a technically feasible way, alleviating the need to hand tune the dose and LET _d distribution. Future work should address multipatient follow‐up studies.

Objective: The purpose of the work presented here was to enable easy access to Monte Carlo dose calculation for both fast neutron therapy and boron neutron capture therapy for research purposes. The dose calculation approach was especially intended to hold high customization potential for individual user applications. Approach: The dose engine is based on the Monte Carlo code MCNP. It is integrated into the MATLAB-based open source research treatment planning software matRad as modular component. Main results: Total dose calculation is enabled for both fast neutron therapy and boron neutron capture therapy on patient CT data. The evaluation of the dose distribution is possible using the matRad graphical user interface and dose volume histograms. Customization options are provided for advanced users. Significance: The open source treatment planning software allows easy access to highly accurate Monte Carlo dose calculation for research projects.

This study presents Fast paRticle thErapy Dose optimizer (FREDopt), a newly developed GPU-accelerated open-source optimization software for simultaneous proton dose and dose-averaged linear energy transfer (LET_d) optimization in intensity-modulated proton therapy treatment planning. FREDopt was implemented entirely in Python, leveraging CuPy for GPU acceleration and incorporating fast Monte Carlo simulations from the FRED code. The treatment plan optimization workflow includes pre-optimization and optimization, the latter equipped with a novel superiorization of feasibility-seeking algorithms. Feasibility-seeking requires finding a point that satisfies prescribed constraints. Superiorization interlaces computational perturbations into iterative feasibility-seeking steps to steer them toward a superior feasible point, replacing the need for costly full-fledged constrained optimization. The method was validated on two treatment plans of patients treated in a clinical proton therapy center, with dose and LET_d distributions compared before and after reoptimization. Simultaneous dose and LET_d optimization using FREDopt led to a substantial reduction of LET_d and (dose) × (LET_d) in organs at risk while preserving target dose conformity. Computational performance evaluation showed execution times of 14–50 min, depending on the algorithm and target volume size—satisfactory for clinical and research applications while enabling further development of the well-tested, documented open-source software.

Background: Robust treatment planning algorithms for intensity modulated proton therapy (IMPT) and intensity modulated radiation therapy (IMRT) allow for uncertainty reduction in the delivered dose distributions through explicit inclusion of error scenarios. Due to the curse of dimensionality, application of such algorithms can easily become computationally prohibitive. Purpose: This work proposes a scenario‐free probabilistic robust optimization algorithm that overcomes both the runtime and memory limitations typical of traditional robustness algorithms. Methods: The scenario‐free approach minimizes cost‐functions evaluated on expected‐dose distributions and total variance. Calculation of these quantities relies on precomputed expected‐dose‐influence and total‐variance‐influence matrices, such that no scenarios need to be stored for optimization. The algorithm is developed within matRad and tested in several optimization configurations for photon and proton irradiation plans. A traditional robust optimization algorithm and a margin‐based approach are used as a reference to benchmark the performance of the scenario‐free algorithm in terms of plan quality, robustness, and computational workload. Results: The implemented scenario‐free approach achieves plan quality similar to traditional robust optimization algorithms, and it reduces the distribution of standard deviation within selected structures when variance reduction objectives are defined. Avoiding the storage of individual scenario information allows for the solution of treatment plan optimization problems, including an arbitrary number of error scenarios. The observed computational time required for optimization is close to a nominal, non‐robust algorithm and substantially lower compared to the traditional robust approach. Estimated gains in relative runtime range from approximately – times with respect to the traditional approach. Conclusion: This work introduces a novel scenario‐free optimization approach relying on the precomputation of probabilistic quantities while preserving compatibility with state‐of‐the‐art uncertainty modeling. The measured runtime and memory footprint are independent of the number of included error scenarios and similar to those of non‐robust margin‐based optimization algorithms, while achieving the required dose and robustness specifications under multiple different optimization conditions. These properties make the scenario‐free approach suitable and beneficial for 3D and 4D robust optimization involving a high number of error scenarios and/or CT phases.

Robust treatment planning algorithms for Intensity Modulated Proton Therapy (IMPT) and Intensity Modulated Radiation Therapy (IMRT) allow for uncertainty reduction in the delivered dose distributions through explicit inclusion of error scenarios. Due to the curse of dimensionality, application of such algorithms can easily become computationally prohibitive. This work proposes a scenario-free probabilistic robust optimization algorithm that overcomes both the runtime and memory limitations typical of traditional robustness algorithms. The scenario-free approach minimizes cost-functions evaluated on expected-dose distributions and total variance. Calculation of these quantities relies on precomputed expected-dose-influence and total-variance-influence matrices, such that no scenarios need to be stored for optimization. The algorithm was developed within matRad and tested in several optimization configurations for photon and proton irradiation plans. A traditional robust optimization algorithm and a margin-based approach are used as a reference to benchmark the performances of the scenario-free algorithm in terms of plan quality, robustness and computational workload. The scenario-free approach achieves plan quality compatible with traditional robust optimization algorithms and it reduces the standard deviation within selected structures when variance reduction objectives are defined. Avoiding the storage of individual scenario information allows for the inclusion of an arbitrary number of error scenarios. The observed optimization time is independent on the number of included scenarios, compatible with a nominal, non-robust algorithm and significantly lower than the traditional robust approach. These properties make the scenario-free approach suitable for the optimization of robust plans involving a high number of error scenarios and CT phases as 4D robust optimization.

Objective: To present a long short-term memory (LSTM)-based prompt gamma (PG) emission prediction method for proton therapy. Approach: Computed tomography (CT) scans of 33 patients with a prostate tumor were included in the dataset. A set of 10⁷ histories proton pencil beam (PB)s was generated for Monte Carlo (MC) dose and PG simulation. For training (20 patients) and validation (3 patients), over 6000 PBs at 150, 175 and 200 MeV were simulated. 3D relative stopping power (RSP), PG and dose cuboids that included the PB were extracted. Three models were trained, validated and tested based on an LSTM-based network: (1) input RSP and output PG, (2) input RSP with dose and output PG (single-energy), and (3) input RSP/dose and output PG (multi-energy). 540 PBs at each of the four energy levels (150, 175, 200, and 125–210 MeV) were simulated across 10 patients to test the three models. The gamma passing rate (2%/2 mm) and PG range shift were evaluated and compared among the three models. Results: The model with input RSP/dose and output PG (multi-energy) showed the best performance in terms of gamma passing rate and range shift metrics. Its mean gamma passing rate of testing PBs of 125–210 MeV was 98.5% and the worst case was 92.8%. Its mean absolute range shift between predicted and MC PGs was 0.15 mm, where the maximum shift was 1.1 mm. The prediction time of our models was within 130 ms per PB. Significance: We developed a sub-second LSTM-based PG emission prediction method. Its accuracy in prostate patients has been confirmed across an extensive range of proton energies.

Objective. To present a long short-term memory (LSTM) network-based dose calculation method for magnetic resonance (MR)-guided proton therapy. Approach. 35 planning computed tomography (CT) images of prostate cancer patients were collected for Monte Carlo (MC) dose calculation under a perpendicular 1.5 T magnetic field. Proton pencil beams (PB) at three energies (150, 175, and 200 MeV) were simulated (7560 PBs at each energy). A 3D relative stopping power cuboid covering the extent of the PB dose was extracted and given as input to the LSTM model, yielding a 3D predicted PB dose. Three single-energy (SE) LSTM models were trained separately on the corresponding 150/175/200 MeV datasets and a multi-energy (ME) LSTM model with an energy embedding layer was trained on either the combined dataset with three energies or a continuous energy (CE) dataset with 1 MeV steps ranging from 125 to 200 MeV. For each model, training and validation involved 25 patients and 10 patients were for testing. Two single field uniform dose prostate treatment plans were optimized and recalculated with MC and the CE model. Results. Test results of all PBs from the three SE models showed a mean gamma passing rate (2%/2 mm, 10% dose cutoff) above 99.9% with an average center-of-mass (COM) discrepancy below 0.4 mm between predicted and simulated trajectories. The ME model showed a mean gamma passing rate exceeding 99.8% and a COM discrepancy of less than 0.5 mm at the three energies. Treatment plan recalculation by the CE model yielded gamma passing rates of 99.6% and 97.9%. The inference time of the models was 9–10 ms per PB. Significance. LSTM models for proton dose calculation in a magnetic field were developed and showed promising accuracy and efficiency for prostate cancer patients.